Model selection for nested model classes with cost constraints

For model selection with nested classes, we propose to minimize Rissanen's stochastic complexity with a constraint on expected computational cost. The proposed solution uses the Wald statistic and dynamic programming for order selection, such that maximum likelihood estimates of only a small subset of hypothesized models need to be computed. Simulation results are presented to compare computational savings and detection performance for superimposed undamped exponentials in additive noise.